Recent headlines have clashed spectacularly on the question of AI and water usage. On the one hand, Sam Altman, CEO of OpenAI (ChatGPT), has suggested that an average AI query uses roughly a fifteenth of a teaspoon of water. On the other, a Morgan Stanley projection estimates that AI data centres could consume around a trillion litres of water per year by 2028.
At face value, these claims seem contradictory. Multiply a teaspoon by any reasonable number of queries and you still do not reach a trillion litres. The uncomfortable truth is that both numbers can be broadly correct and still mislead.
The problem with “per-query” thinking
A single user prompt is rarely a single computation. Modern AI systems interpret prompts, perform safety checks, run internal reasoning steps, sometimes query other models or tools, and then evaluate outputs before returning an answer. What looks like one query to a user may involve many unseen computational steps, each with energy and cooling requirements.
That said, even if a query triggers multiple internal steps, the water used at the moment of interaction remains small. A few teaspoons of water per response is not, on its own, the global problem. The real issue lies elsewhere.
What per-query numbers leave out
Per-query figures typically measure only the final inference step when you interact with a chatbot. They exclude the most resource-intensive parts of AI’s lifecycle, particularly training.
Training large models runs for weeks or months on vast clusters of GPUs, consuming enormous amounts of electricity and water for cooling. Crucially, this training is not a one-off event. Even while a model is in use, companies are continuously training newer, larger systems. From a resource-accounting perspective, it is difficult to justify treating training as separate from everyday use: without it, the queries could not exist.
How you distribute that training footprint across user interactions is a choice and different choices produce wildly different numbers. Spread it narrowly and water use looks tiny; spread it across the system’s lifetime and it grows quickly.
When big numbers overstate the case
At the opposite extreme, some analyses inflate AI’s water footprint by counting all water withdrawn by power plants supplying electricity to data centres. Thermoelectric power generation does indeed move vast quantities of water, but most of it is returned to rivers or lakes, with only a small fraction lost to evaporation. Including total withdrawals can make AI appear far more water-intensive than it truly is, especially when compared with municipal drinking-water use.
This does not mean power-plant water is irrelevant. Waste heat and evaporation still matter but not all water is is treated the same, so lumping categories together obscures more than it clarifies.
Water is a local problem, not a global average
The most important insight is that water impact is context-dependent. A water-cooled data centre in a desert competes directly with already-stressed ecosystems and communities. The same facility in a water-abundant region may pose fewer risks. Global averages hide this reality. What matters is not just how much water is used, but where, what kind, and under what local constraints.
Perspective matters
Even under ambitious growth scenarios, AI’s water use remains small compared with other industries we rarely question. Agriculture, particularly irrigated crops grown for fuel rather than food, consumes orders of magnitude more water than global AI infrastructure. That does not excuse poor AI planning, but it does highlight how selectively we frame “waste” and “necessity.”
The real takeaway
AI’s water footprint is neither negligible nor apocalyptic. It is complex, poorly disclosed, and easy to misrepresent in either direction. Small per-query figures create false reassurance; while annual totals can exaggerate impacts by blurring categories.
The honest conclusion is simpler and less comfortable: we do not yet have the transparency needed to account for AI’s water use properly. Until we do, the most responsible focus is on location selection, system design, water recycling, and perhaps more importantly the rapidly growing energy demand that underpins AI.
As with carbon, what matters most is not the headline number, but whether we are asking the right questions in the right places. Sustainable use of resources needs to be a key priority in the expansion AI capabilities and it is not always clear that it is.